EUROGRAPHICS 2008 / G. Drettakis and R. Scopigno
(Guest Editors)
Volume 27 (2008), Number 2
Effect of Character Animacy and Preparatory Motion on
Perceptual Magnitude of Errors in Ballistic Motion
P. S. A. Reitsma1 , J. Andrews2 , and N. S. Pollard3
1 Trinity
College Dublin, Ireland
of California, Berkeley, USA
3 Carnegie Mellon University, USA
2 University
Abstract
An increasing number of projects have examined the perceptual magnitude of visible artifacts in animated motion.
These studies have been performed using a mix of character types, from detailed human models to abstract geometric objects such as spheres. We explore the extent to which character morphology influences user sensitivity
to errors in a fixed set of ballistic motions replicated on three different character types. We find user sensitivity
responds to changes in error type or magnitude in a similar manner regardless of character type, but that users
display a higher sensitivity to some types of errors when these errors are displayed on more human-like characters. Further investigation of those error types suggests that being able to observe a period of preparatory motion
before the onset of ballistic motion may be important. However, we found no evidence to suggest that a mismatch
between the preparatory phase and the resulting ballistic motion was responsible for the higher sensitivity to
errors that was observed for the most humanlike character.
Categories and Subject Descriptors (according to ACM CCS): I.3.7 [Computer Graphics]: Three-Dimensional
Graphics and Realism
1. Introduction
A substantial body of recent work has examined the problem of quantifying the perceptual magnitude of errors in animated motion. While data on a variety of errors and scenarios has been generated, these results come from studies
(a)
(b)
(c)
(d)
Figure 1: A selection of the character types used in recent perceptual studies of animation. (a) O’Sullivan et al.
[ODGK03] (b) Reitsma and Pollard [RP03] (c) Harrison et
al. [HRvdP04] (d) O’Sullivan and Dingliana [OD01]
c 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation Published by Blackwell Publishing, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350
Main Street, Malden, MA 02148, USA.
using a wide array of different character types, from human
figures of varying realism to abstract geometric objects such
as circles (Figure 1). Unfortunately, the character differences
make comparing results between these studies very difficult,
as the interaction between character type and user sensitivity
to error is largely unknown.
There is some evidence that improved graphical quality of
animations may increase the ability of users to detect anomalous motions (e.g., [SW93] [HOT98] [OHJ00] [HB00]). In
particular, Hodgins and her colleagues [HOT98] demonstrated that subjects are more sensitive to motion changes
displayed on polygonal humanoid characters as compared to
humanoid characters formed by stick figures. The reason for
this difference is unknown, however; one hypothesis is that
the more realistic polygonal character allows greater sensitivity than the simpler and more abstract stick figure.
We explore this hypothesis, considering the extreme difference between a human character and a sphere. In particular, we ask whether the findings of previous researchers are
P. S. A. Reitsma, J. Andrews, & N. S. Pollard / Effect of Char. Animacy and Prep. Motion on Percep. Magnitude of Errors in Ballistic Motion
Figure 2: An example of one of the ballistic motions used in our user studies, depicted on two of the characters used: a human
figure (top) and a cannonball (bottom) which followed the same center of mass trajectory. The second frame is the takeoff point
where the character transitions into ballistic motion.
the entire flight phase. We present full human body motion
through a human character, and identical center of mass motion through a spherical object (Figure 3). Our primary motivation is to answer three questions:
Figure 3: The set of characters employed in our user studies. Left two characters were used in the first study; right two
characters were used in the second study.
replicated or contradicted in the case of ballistic motion.
Consider a human character and a sphere, both of which
undergo ballistic motion with identical center of mass trajectories (Figure 2). Intuitively, one might expect that errors
in the ballistic trajectory would be easier to detect with the
ball “character”, due to the simplicity of its motion and the
lack of distractions such as limb motion. Our familiarity with
human motion, though, makes humans very skillful at detecting anomalies in human motion, suggesting that perhaps
subjects would draw on their experience to display greater
sensitivity to errors in the trajectory of the human character. Without some knowledge of how sensitivity to errors
varies with character type, our ability to compare and exploit the results of studies performed on different characters
is limited. Indeed, neuroimaging studies (e.g., [GDP∗ 00]
[PMM∗ 03]) suggest that different regions of the brain are
active when observing coherent motion deemed biological
or non-biological, suggesting that there may be fundamental
and irreconcilable differences between user perceptions of
animated motion in complex human characters and in simple, abstract objects such as spheres.
In this paper, we explore ballistic motions derived from
human motion captured jumps. We add errors of three varieties, motivated by the types of errors that arise during motion editing and transitions between motions: a quick change
in vertical velocity during the flight phase; a quick change
in horizontal velocity (in the direction of the jump) during the flight phase; and a change in apparent gravity over
1. Does a difference in user ability to detect added errors
exist between realistic human characters and simple, abstract characters for the particular case of ballistic motion?
2. If a difference does exist, is it consistent? i.e., are the perceptual characteristics of motion on one character type
reliably predictable from the perceptual characteristics of
motion on the other character type?
3. If there is a consistent difference, how much and in what
direction? i.e., are changes to ballistic motion easier to
detect on a simple, abstract character on or a realistic human character?
Additionally, we examine potential underlying causes for
the observed differences. In particular, we explore the effect
of displaying a preparatory motion for the ball, as shown in
Figure 6.
In summary, we find no difference in user ability to detect horizontal or vertical velocity errors regardless of character type, but we find that changes in the level of gravity
are easier to detect with a human character than with a ball
character. The sensitivity difference was smaller if a preparatory motion for the ball was displayed then when it was not;
however, subjects were significantly more sensitive to motion of the human character in both cases. Perhaps surprisingly, there was no evidence to suggest that sensitivity was
affected by a mismatch (a different gravitational constant)
between the preparatory motion and the ballistic motion of
the ball. In other words, it did not appear to matter whether
the preparatory motion was consistent with the ballistic motion, only that a preparatory motion was present.
2. Background
A number of researchers have examined the perceptual magnitude of errors in motion. Most have examined the phenomenon using geometric objects such as circles or dots
c 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation P. S. A. Reitsma, J. Andrews, & N. S. Pollard / Effect of Char. Animacy and Prep. Motion on Percep. Magnitude of Errors in Ballistic Motion
(e.g., [Mic63] [Coh64] [Run74] [SW93] [OD01]), spheres
and rigid bodies [ODGK03], or two-link chains [HRvdP04],
although recently some researchers have looked at human
characters [RP03] [WB03] [WB04] [RPE∗ 05].
Researchers have also examined the effects on user perception of differences in character model or animation. Hodgins et al. [HOT98] showed that subjects are more sensitive
to differences between pairs of motions when those motions
are displayed on a polygonal character rather than a stick
figure character; similarly, Chaminade et al. [CHK07] found
that subject ability to discriminate between biological and
non-biological motion was impaired when that motion was
displayed on a character that was a cloud of points rather
than a polygonal model character or even a character built
up out of ellipses. Stappers and Waller [SW93] noted that
more complex visual stimuli, such as more droplets in an
animated fountain, made users more sensitive to depth estimation in a virtual scene. Oesker et al. [OHJ00] showed that
greater detail and realism in the animation of soccer-playing
humanoids allowed subjects to more accurately gauge the
skill with which the humanoids were playing. All of these
experiments indicate that character complexity can affect our
ability to make judgements about motion; however, the effect of character differences on user sensitivity to physical
errors in character motion has not yet been examined.
It is generally acknowledged that people perform poorly
on abstract physical reasoning tasks [Pro99], although animation has been shown to improve performance [KPWH92].
Hecht and Bertamini [HB00] showed that subjects have a
poor innate understanding of ballistic motion and were tolerant of substantial errors in ballistic trajectory for animations of thrown balls, including added accelerations and
decelerations. Neuroimaging studies (e.g., [GDP∗ 00] and
[PMM∗ 03]), however, make it unclear whether the findings
of Hecht et al. on simple projectiles apply to recognizablyhuman characters, although Chaminade et al. [CHK07] do
not find a consistent difference in brain activity between
character types.
Reitsma and Pollard [RP03] examined user sensitivity to
errors in the ballistic phase of motion captured human jumping motions. We adopt identical techniques to examine the
sensitivity differences of subjects to errors in motions displayed on rigid bodies vs. humanoid characters. As with
their approach, we use detection theory [MC91] to convert
rating data to sensitivity measures.
3. Study 1: Effect of Character Animacy on Error
Sensitivity
Our first study was designed to answer our three core questions:
1. Are errors in ballistic motion more apparent with realistic
human characters or with simple sphere "characters"?
2. If a difference does exist, is it consistent?
c 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation 3. If there is a consistent difference, how much and in what
direction?
We selected the experimental methodology of Reitsma
and Pollard [RP03] for our experiment; one benefit of this
choice was the ability to compare our two character models
(realistic and abstract; see Figure 3) to their semi-realistic
character model (Figure 1(b)), enhancing the discriminatory
power of our experiment.
Our two characters, a simple ball and a human model,
were chosen to represent two extremes in the range of possible characters, the ball being one of the simplest possible
characters, with no moving parts, and the human being a
complex, but intuitively understood, character.
Participants: Participants were obtained by university-wide
advertising. 9 women and 13 men ranging in age from 18 to
33 successfully completed the study; four additional participants did not follow instructions and were excluded from the
analysis.
Participants were excluded based on whether their responses on the exit questionnaire clearly revealed that they
had not followed instructions. The most common reason for
exclusion was a failure to follow the instruction that animations were to be judged based on their appearance while the
character was in the air; for example, one participant was
excluded for indicating that they based some of their decisions on how the human character behaved after finishing
the landing phase of its jump.
Stimuli: Participants were shown animations of two characters – a textured human character and a spherical cannonball
– undergoing ballistic motion. All animations were shown in
the same rendering style, with the same (fixed) camera configuration (Figure 2), with the character beginning the motion at the same position and jumping in the same direction
each time. Shadows were rendered, and a small amount of
motion blur was added. These parameters were chosen so as
to make this user study as comparable as possible to the studies performed by Reitsma and Pollard [RP03], as were other
details of our experimental setup, such as user study conditions, participant instructions, camera angle, and apparent
character size.
Animations of human jumping motions were created as
stimuli using seven source motions, which were modified
with three error varieties and three error magnitudes as listed
in Table 1. Horizontal and vertical errors were created by
smoothly adding the indicated level of velocity to the root
of the character over a 0.1s time window starting 0.1s after
the beginning of the ballistic phase of motion. Gravity errors
were created by altering the level of gravity over the entire
ballistic phase of the motion. For motions whose ballistic
phase changed duration due to the added error, the non-root
motion of the character was timewarped to fit the new duration of the ballistic phase, with linear interpolation between
P. S. A. Reitsma, J. Andrews, & N. S. Pollard / Effect of Char. Animacy and Prep. Motion on Percep. Magnitude of Errors in Ballistic Motion
Error
Variety
Horizontal
Vertical
Gravity
Small
± 0.45m/s
± 0.45m/s
± 1.70m/s2
Error Magnitude
Medium
Large
± 0.73m/s
± 1.10m/s
± 0.73m/s
± 1.10m/s
± 2.70m/s2 ± 4.00m/s2
Table 1: Error magnitudes used for our first user study.
values at adjacent frames used to resample joint poses as
needed. In all cases, character root velocity at the end of
the ballistic phase was edited via displacement splines to
match the root landing velocity in the original motion capture. These displacement splines were extended as far back
into the ballistic motion as possible to avoid presenting a
second source of perceptible errors. Source motions and error treatments were identical to those used in Reitsma and
Pollard [RP03].
Motions for the ball character consisted of three parts
linked together in a C1-continuous manner, corresponding
to the takeoff, ballistic, and landing phases of the jumps performed by the human character. The takeoff phase for the
ball consisted of smooth linear acceleration from a fullyhidden position inside the cannon’s barrel to the point where
the ball was half-emerged from the mouth of the cannon
(first and second frame respectively of Figure 2). The ball
then entered its ballistic phase, and followed the center of
mass trajectory of the human character performing that motion (frames three through five of Figure 2). Finally, the
frame where the human character’s feet touch the ground
marks the beginning of the landing phase; during this phase,
the ballistic motion of the ball was extended by continuing
its motion at constant gravitational acceleration until it entered a basket placed on the ground plane and disappeared
from view (last two frames of Figure 2). In order to provide scale cues, the human character was placed beside the
cannon, and remained motionless throughout the animation
(Figure 3).
Procedure. Participants were told they would see two types
of animated motions: a human jumping, or a ball traveling
through similar arcs. They were given some background information on how motion capture data is created. Participants were told that half of the motions had errors, that errors were similar in each of the two types of animations, and
that errors occurred during the flight phase of a motion.
Participants were then shown a training set of 24 motions,
consisting of each of the two characters (human and ball)
being used to display the same 12 representative animations.
They were told that half contained errors but were not told
which specific motions in this training set contained errors.
Source motions for each error treatment were chosen at random, and the presentation order of the full set of 24 animations was randomly permuted. This training set allowed
users to see motions with errors of all detectability levels, allowing them to calibrate their use of the 0–9 scale and make
use of the full range of available responses. Not only did
this help prevent a learning effect from skewing the initial
responses, using the full range of the scale provides greater
separation of the data, and hence more information on the
sensitivity of the participant.
Each error treatment was shown on two different source
motions, for 36 unique error-containing trials. The original
source motion for each of these trials was also included,
bringing the total to 72 motions. Finally, each motion was
displayed on each of the two characters, for a total of 144
test motions. The order of presentation of the 144 animations
was randomly permuted, and then split into four blocks of 36
trials for display purposes.
All motions were placed on DVD in movie format and
played on a commercial DVD player, as that appeared to
minimize playback hitches compared to playback from either a video file on computer or from a DVD in a computer
DVD drive. Six DVDs were used, each with a different random assignment of source motions to error treatments, and
each with a unique permutation of presentation order for the
resulting animations. Each DVD was seen by between 3 and
5 subjects.
Stimuli were presented on a projection screen in a small
conference room. Participants were instructed to categorize
each motion as either “no error (unchanged)” or “error” and
mark their level of confidence in their answer using a rating
scale that ranged from 0 (most confident an error is present)
through 9 (most confident an error is not present). Participants categorized training motions as well as test motions,
but the data from the training motions was not used for any
analyses.
At the end of the study, participants were asked to describe
their experience in the study of motion, including involvement in sports, dance, video games, etc.
3.1. Detection Theory
As motion capture data provides only a lossy approximation
of the original human motion, all user judgements will be inherently subjective. Accordingly, substantial response bias is
possible; for example, one respondent might think all motion
captured data looks poor, and hence rate all motions with a
low score, whereas another respondent might like animated
motion and tend to rate all motions highly. Unless taken into
account by the analysis technique, such between-subject differences would dilute the findings of the experiment.
As noted by Reitsma and Pollard [RP03], detection theory [MC91] can be used to derive a bias-independent measure of a user’s ability to detect errors in an animated motion.
The method takes into account the difference between how
frequently the subject correctly labelled a motion as containc 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation P. S. A. Reitsma, J. Andrews, & N. S. Pollard / Effect of Char. Animacy and Prep. Motion on Percep. Magnitude of Errors in Ballistic Motion
ing an error (hit rate H) and how frequently the subject incorrectly labelled an unchanged motion as containing an error
(false alarm rate F). For the simplest experimental design (a
yes/no response), a subject’s sensitivity (d) to errors would
be computed as:
d = z(H) − z(F)
(1)
where z is the inverse of the normal distribution function.
For example, a hit rate of 75% and a false alarm rate of 25%
corresponds to a sensitivity of 1.35, as does a hit rate of 90%
coupled with a false alarm rate of just over 47%. These two
examples of how to obtain a sensitivity of 1.35 demonstrate
the bias-independent nature of detection theory: as sensitivity is computed based on the relative distribution of the participant’s responses rather than on the raw distribution, factors which will systematically bias the responses, such as
participant reaction to the quality of the animation, are automatically factored out.
A similar approach allows sensitivity levels to be computed from rating data (intuitively, a high mean rating corresponds to a high hit rate and high false alarm rate; see
[MC91] for details).
3.2. Study 1 Results
Figure 4 shows mean sensitivities for all error varieties and
directions for both characters used in our first user study. The
plots show the mean sensitivities of motions at each magnitude of added error, including unchanged motions. Mean
values from Reitsma and Pollard’s 2003 paper [RP03] are
plotted for comparison, and are directly comparable.
We used the residual maximum likelihood (REML)
method (see [PT71] [CS76]) to analyze the per-subject sensitivity values (computed as per [MC91]) with 3 error levels
x 2 character types (human and ball) x 3 error varieties x 2
error directions. This technique was chosen due to our use of
a mixed model with subject ID as a random effect; this design allowed inter-subject variability to be taken into account
more explicitly. All error varieties could be detected with
P < 0.01, and no effect of experience (F(1, 736) = 1.88, P =
0.17) or gender (F(1, 736) = 0.56, P = 0.46) was found.
We found that error type affected perceptual differences
between different characters:
(1) Subjects found gravity errors easier to detect with human characters than with ball characters
F(1, 233) = 5.98, P = 0.015
(2) Subjects displayed no significant difference in ability
to detect either horizontal or vertical errors regardless of
whether human characters or ball characters were used
F(1, 233) = 2.03, P = 0.16 and F(1, 233) = 0.002, P = 0.97,
respectively
Our study gave very comparable overall results to the
c 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation studies reported by Reitsma and Pollard [RP03]. Our study
confirmed the three main effects found in their work:
(RP1) Subjects found added acceleration easier to detect
than added deceleration
Human: F(1, 238) = 44.33, P = 0.001
Ball: F(1, 238) = 15.85, P < 0.001
(RP2) Subjects found low gravity easier to detect than
high gravity
Human: F(1, 107) = 32.48, P < 0.001
Ball: F(1, 107) = 13.09, P < 0.001
(RP3) Subjects found errors in horizontal velocities easier to detect than errors in vertical velocities
Human: F(1, 238) = 20.13, P < 0.001
Ball: F(1, 238) = 25.57, P < 0.001
Sensitivity to horizontal and vertical errors did not differ significantly between the studies (F(1, 53) = 1.64, P =
0.20). Sensitivity to gravity errors did not differ significantly
between the previous study and our study’s human character
(F(1, 31) = 3.58, P = 0.068) or between the previous study
and our study’s ball character (F(1, 31) = 3.82, P = 0.060).
This result is especially interesting in light of the substantial difference between our human and ball characters (main
effect 1); one possible explanation is that perhaps the rough
human figure used by Reitsma and Pollard may have been
“perceptually in between” our realistic human character and
simple ball character.
4. Study 2: Effect of Preparatory Motion
Study 1 identified no significant difference in user sensitivity between character types for the case of vertical or horizontal velocity perturbations, but did identify a substantial
difference for the case of changes to gravity over the entire ballistic phase of the motion (see Section 3.2). The goal
of our second study was to investigate possible mechanisms
responsible for this difference. In particular, our intent was
to control for possible sources of systematic bias in the experimental design to narrow down what factors caused the
observed difference. Possible sources of systematic bias include:
1. Mismatch between preparatory and ballistic motion only
for the human character.
2. Signalling of landing position only for the ball character.
3. Inherent scale cues only for the human character.
In Study 2, we correct these three differences and further
investigate the first of these differences as a possible cause
of the results in Study 1.
In Study 1, human character animation contained more
information in the form of preparatory motion. In particular, the “apparent effort” of a human preparing for a jump
could be observed in the human motions, whereas the ball
P. S. A. Reitsma, J. Andrews, & N. S. Pollard / Effect of Char. Animacy and Prep. Motion on Percep. Magnitude of Errors in Ballistic Motion
Figure 4: Mean sensitivities for all errors from both characters used in the first study, as well as the character from Reitsma
and Pollard [RP03]. Error bars show standard error of the mean.
Figure 5: An example of the preparatory and ballistic motion seen in our second user study. The character transitions to
ballistic motion just prior to frame 3.
displayed no “preparatory motion”, as it was launched from
a cannon. A mismatch between the perceived effort and the
actual height or distance of a jump was hypothesized to explain the difference in sensitivity.
using a new ball character with the size and textured appearance of a basketball (see Figure 6). As with the first experiment, the human character was positioned beside the launch
apparatus to provide additional scale cues.
To test this hypothesis, in the new study, the ball character was given preparatory motion; i.e., shown rolling along
a launch ramp before undergoing ballistic motion (Figure
5). Though a number of different launch devices, such as a
spring, a rubber band, or a ‘pushing rod,’ could be used to
explain the ball’s launch velocity, rolling the ball down the
ramp without introducing any external launch devices was
chosen as the clearest, simplest method for generating the
preparatory motion. In particular, it did not introduce any
additional moving parts to the animation, and the correct effect of gravity may be intuitively understood.
Participants: Participants were obtained by university-wide
advertising. 9 women and 10 men ranging in age from 18
to 32 successfully completed the study; three additional subjects were excluded from the analysis due to hardware failure during their session.
To remove the second identified source of bias, both human and ball characters had their landing points marked in
a similar manner (a manhole cover for the human to land on
and an equivalent-sized hole for the ball to vanish into). Finally, we addressed the question of inherent scale cues by
Stimuli: Participants were shown animations of two characters – a textured human character and a textured spherical ball – undergoing preparatory and then ballistic motion.
The experimental setup was similar to the previous study,
but was changed to accommodate the launch ramp and larger
error magnitudes (Table 2); accordingly, character textures,
camera angle, and camera distance from the characters had
slightly different (fixed) values.
This study used the same source ballistic motions as the
previous study, and applied gravity errors in the same manc 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation P. S. A. Reitsma, J. Andrews, & N. S. Pollard / Effect of Char. Animacy and Prep. Motion on Percep. Magnitude of Errors in Ballistic Motion
Figure 6: Experimental setup used to display preparatory
motion for the ball character in the second study. The ball
rolls along the launch ramp under either the correct gravitational force or under a gravitational force consistent with
the following ballistic motion.
Error
Direction
Decreased
Increased
Error Magnitude
Small
Medium
Large
2
2
± 2.7m/s
± 4.0m/s
± 5.5m/s2
± 3.4m/s2 ± 5.4m/s2 ± 8.0m/s2
Table 2: Gravity error magnitudes used for our second user
study. Magnitudes were based on pretesting with the new experimental setup.
Figure 7: Mean sensitivities for all character types used
in our second study. Error bars show standard error of the
mean.
ner; prior to undergoing ballistic motion, three cases of
preparatory motion were used:
• Correct: the ball accelerated according to normal gravity
(i.e., 9.8m/s2 ).
• Consistent: the ball accelerated according to the level of
gravity present in the ballistic motion.
• Hidden: the ball could not be seen prior to ballistic motion.
of 36 unique motions – 3 examples of each error treatment
shown on three different source motions, plus the corresponding error-free motions – displayed on each of the four
characters, for a total of 144 test motions. The order of presentation of the 144 animations was randomly permuted, and
then split into four blocks of 36 trials for display purposes.
The Correct case corresponds to the situation for the human
character: since the human is animated using the untouched
motion capture data during the takeoff phase of the jump, it
is subject to the correct level of gravity during that takeoff,
resulting in an acceleration mismatch between preparatory
and ballistic motion. By contrast, the Consistent case has no
mismatch between the level of gravity experienced by the
ball during preparatory and ballistic motion; gravity is consistent through all phases of the motion. Finally, the Hidden
case corresponds to the cannonball character from the first
study, as the preparatory motion can not be seen.
Four DVDs were used, each with a different random assignment of source motions to error treatments, and each
with a unique permutation of presentation order for the resulting animations. Each DVD was seen by between 4 and 5
subjects.
Procedure:
Participants were given similar instructions to the previous study, were shown 20 representative training examples,
and then were shown 144 test motions. The experimental apparatus (projection screen, DVD playback, response sheets,
etc.) was identical to the previous study. Each test consisted
c 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation 4.1. Study 2 Results
Figure 7 shows mean sensitivities for all character types and
both directions used in our second user study. Results are
broken out by small, medium, and large error levels.
As with the first experiment, we used the REML method
to analyze the per-subject sensitivity values, with 3 error levels x 4 character treatments x 2 error directions. All error varieties could be detected with P < 0.01 except for increased
gravity errors in the Hidden case, and no effect of experience (F(1, 418) = 0.25, P = 0.62) or gender (F(1, 418) =
0.21, P = 0.65) was found.
P. S. A. Reitsma, J. Andrews, & N. S. Pollard / Effect of Char. Animacy and Prep. Motion on Percep. Magnitude of Errors in Ballistic Motion
The second study produced three main effects:
(1) Subjects found gravity errors easier to detect when
displayed on a human character with preparatory motion than when displayed on a ball character with
preparatory motion.
F(1, 426) = 5.58, P = 0.019.
(2) Subjects found increased gravity errors displayed on
ball characters easier to detect when shown preparatory
motion than when they were shown no preparatory motion
F(1, 336) = 6.01, P = 0.015.
(3) Subjects showed no significant difference in sensitivity when preparatory motion for the ball character was
computed with normal gravity or with gravity consistent
with the (altered) ballistic phase of the motion.
F(1, 222) = 0.41, P = 0.52.
In addition to the first main effect, relating the Consistent and Correct cases of the ball character to the human
case, subjects also found errors easier to detect on the human
than on the ball in the Hidden case (F(1, 227) = 18.94, P <
0.001). Sensitivity to the human character was significantly
different from sensitivity to any case of the ball character
(Student’s t = 1.97, α = 0.05). Sensitivity was significantly
higher to decreased gravity on the human character than on
any other condition, and was significantly lower to increased
gravity in the Hidden case than for any other condition, but
there was insufficient data to further discriminate between
{Type x Direction} conditions.
5. Discussion
In these studies, we measured sensitivity of human subjects
to errors in animated ballistic human and ball motion. We
found that changes in the level of gravity were easier to detect with a human character than with a ball character, but
that there was no significant difference in sensitivity between
human characters and ball characters for horizontal or vertical errors. Similarly, we found that changes in the level of
gravity displayed on a ball character were easier to detect
when shown preparatory motion, but that the correctness of
the preparatory motion did not make a significant difference,
and showing the preparatory motion did not erase the sensitivity difference between human and ball characters.
Based on these results, we can address our motivating
questions from Section 1:
1. Does a difference exist?
2. Is the difference consistent?
3. In which direction is the difference?
Our results indicate that perceptual differences exist between the sensitivity of users to changes in motions displayed on abstract ball or realistic human characters, but
only for some types of changes. For our case of ballistic
motion, the sensitivity of users to local perturbations in the
horizontal or vertical center of mass velocity did not differ
between the characters tested; i.e., the difference was consistent, and approximately zero. This result stands in apparent conflict with the results of Hodgins et al. [HOT98] and
Chaminade et al. [CHK07], and further research is needed to
understand under what circumstances character type will affect our ability to judge the realism of motion. One possible
explanation is that simple and abrupt changes to the character’s root motion, as may result from splicing and editing
operations, may be less influenced by character type than
the more complex motion modifications explored in those
two prior studies. Indeed, anecdotal evidence from our poststudy questionnaire suggested that subjects tended to detect
horizontal and vertical errors via direct observation of their
effects, whereas gravity errors, which had less abrupt and
more complex effects, were detected via observation of their
effect on the overall character of the motion.
Changes in the level of gravity did result in significantly
different levels of sensitivity for the different characters.
This difference in user sensitivity between characters was
consistent for decreases in gravity, with subjects showing
approximately twice the sensitivity to decreased gravity displayed on the human character as on the ball character. By
contrast, the sensitivity difference between characters was
not consistent for increases in gravity, with sensitivity on the
more realistic character appearing to rise much more rapidly
with large error magnitudes.
Our second study examined the differences in sensitivity
to altered gravity in more detail, removing several possible
sources of bias and testing the hypothesis that the higher sensitivity found in the first study for the human character is
explained by a mismatch between preparatory motion and
subsequent ballistic motion; in essence, that the apparent effort of the character was often too much or too little for the
magnitude of the following jump.
Our results showed that the presence and correctness of
preparatory motion made no difference for user sensitivity to decreased gravity displayed on the ball character. All
cases of preparatory motion resulted in perceptual magnitudes which were lower than for the human character by
roughly 30%; however, the sensitivity difference varied substantially from that mean difference in the case where the
preparatory motion was hidden.
By contrast, each preparatory motion treatment of the ball
character resulted in qualitatively different user sensitivities
for increased gravity errors. The ball displayed with preparatory motion consistent with its ballistic motion showed the
same pattern as before: user sensitivity consistently lower
than the human character by roughly 30%. However, despite
the lack of significant difference in sensitivity between the
two types of preparatory motion, there was a qualitative difference for increased gravity. User sensitivity to the ball disc 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation P. S. A. Reitsma, J. Andrews, & N. S. Pollard / Effect of Char. Animacy and Prep. Motion on Percep. Magnitude of Errors in Ballistic Motion
played with “correct” preparatory motion (i.e., whose gravity was always 9.8m/s2 regardless of gravity during the ballistic phase) resulted in perceptual magnitudes for the gravity
increases which did not differ by a consistent amount from
the results for the human character, but which were qualitatively similar to those for the cannonball in the first study:
i.e., relatively high sensitivity to medium errors as compared
to sensitivity to small and large errors.
Finally, users had no sensitivity to increased gravity errors displayed on the ball character whose preparatory motion was hidden. One likely factor contributing to this lack of
sensitivity is that motions with increased gravity and hidden
preparatory motion were visible for only a brief time; when
gravity was increased by the maximum amount, the longest
of these motions was visible for 20% less time than the shortest motion for which preparatory motion was shown. There
is no indication from the results, however, that difference
in motion duration between trials within a single error treatment has an effect on the score assigned by subjects. The raw
source motions had flight phase durations which differed by
up to 20%, but no effect of flight phase duration was found
among the scores given to error-free motions in the second
study as a whole (F(1, 2718) = 1.40, P = 0.24), or to motions whose preparatory motion was hidden (F(1, 679) =
0.01, P = 0.92). Accordingly, it appears as though raw duration of a motion may not fully explain the low user sensitivity to errors in cases with hidden preparatory motion.
In all cases, user sensitivity to gravity errors displayed
on the human character was significantly higher than user
sensitivity to errors displayed on the ball character, suggesting factors other than those controlled for in the second experiment must account for a significant portion of the difference. One possibility is that because gravity errors correspond to incorrect behavior over the entire jump, rather
than a localized disruption, subjects’ greater familiarity with
human jumping motion than ball trajectories would account
for the difference, especially because it is known that humans have a poor intuitive sense for the physics of the ballistic trajectories of simple objects such as balls [HB00].
Another possibility is that because a change in the force
of gravity corresponds to timescaling of the jump, the animated motions of the limbs and head of the human character offered additional information that was not present in the
ball character. Stappers and Waller [SW93] note that richer
stimuli improved accuracy and reliability in using the free
fall of objects under gravity to estimate visual depth, suggesting that the additional information offered by the limbs
of the human character may have improved user results in
our experiment. They also note how observed gravity can
vary with perceived scale and distance, suggesting the possibility that the human jumping innately embodies a sense
of scale, whereas subjects may have some freedom to interpret the size and distance of the cannonball to suit the
observed motion, notwithstanding the human figure placed
beside the cannon and the use of a clearly identified basketc 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation ball for the spherical object in the second study. Finally, it is
worth noting that observing biological motion engages our
mirror neurons [GDP∗ 00] [PMM∗ 03] – allowing us to internally mimic an observed motion – in a way non-biological
motion does not. One speculative but interesting possibility
is that this neurological difference makes humans inherently
more sensitive to certain types of errors in human motion
than in motion perceived as non-biological or inanimate.
6. Conclusions
This work examined the extent to which the animacy and
human-like nature of a character affects user sensitivity to errors in ballistic motion, and the effect of displaying matched,
mismatched, or no preparatory motion for simple characters.
Our experiments demonstrated that user sensitivity to errors in ballistic motion follows the same general linear pattern for characters ranging between a human and a ball, suggesting that it is not unreasonable to attempt to generalize
results and findings from one character type to another. Substantial care must be taken with such generalizations, however, as our experiments also demonstrated that for some
types of errors subjects tend to be more sensitive to motion
displayed on human characters than on rough human figures
or simple geometric objects. Whether or not preparatory motion matches ballistic motion appears to play a limited role
in this difference. As ballistic motion is substantially simpler
than general motion, due to the lack of ground contacts, this
result suggests animators may not need to aggressively extend edits in ballistic motion into the ground contact phases
before and after, potentially saving substantial effort.
Additionally, the low user sensitivity to increased gravity
in the case with hidden preparatory motion may indicate that
fast, smooth motions which are visible for only a short time
may not allow a user to fully evaluate their physical plausibility, and hence may look acceptable even with substantially larger errors. As rapid motion out of concealment is a
common scenario is applications such as real-time computer
games, we are interested in further exploring the extent to
which brief, unexpected, or partially-occluded motions may
be able to absorb higher levels of error for the same loss of
perceptual quality than slower, longer, and more expected
motions.
Finally, we note that the error types for which character
morphology made no difference to user sensitivity (added
velocity in the horizontal or vertical direction) are inherently local changes to the motion; subject responses to the
post-study questionnaire suggested that these types of motions were typically detected by direct observation of their
effects (e.g., the character was “pushed”, or the motion was
jerky). By contrast, the errors for which user sensitivity differed significantly between the character types (increased
or decreased gravity) are global changes which extend over
a much wider duration of the underlying motion, and sub-
P. S. A. Reitsma, J. Andrews, & N. S. Pollard / Effect of Char. Animacy and Prep. Motion on Percep. Magnitude of Errors in Ballistic Motion
ject responses suggested that these types of errors were typically detected by indirect observation (e.g., the character
“floated”, or travelled “too far”). Intuitively, we would expect a subject’s greater familiarity with human motion to
assist their detection when motions are detected indirectly
by their effect on the overall character of the motion, but to
have a more limited effect when errors are detected by direct observation. We are interested in determining whether
this local/global split is an artifact of the particular errors we
tested, or whether it might apply more generally.
7. Acknowledgements
Supported in part by the IRCSET Embark Initiative and
by the NSF under grants ISS-0326322, CCF-0343161, and
ECS-0325383. Alias/Wavefront donated their Maya software for use in this research. The authors would like to thank
Raphael Mun, Daniel Hill, and Moshe Mahler for their assistance with this work.
References
[CHK07] C HAMINADE T., H ODGINS J., K AWATO M.:
Anthropomorphism influences perception of computeranimated characters’ actions. Social Cognitive and Affective Neuroscience, 2(3); p.206-216 (2007).
[Coh64] C OHEN R. L.: Problems in motion perception.
Uppsals, Sweden: Lundequistska (1964).
[CS76] C ORBEIL R. R., S EARLE S. R.: Restricted maximum likelihood (reml) estimation of variance components
in the mixed model. Technometrics, Vol 18., No. 1 (1976).
[GDP∗ 00] G ROSSMAN E., D ONNELLY M., P RICE R.,
P ICKENS D., M ORGAN V., N EIGHBOR G., B LAKE R.:
Brain areas involved in perception of biological motion.
Journal of Cognitive Neuroscience, 12:5, pp. 711-720
(2000).
[HB00] H ECHT H., B ERTAMINI M.: Understanding projectile acceleration. Journal of Experimental Psychology:
Human Perception and Performance 26, 2 (2000), 730–
746.
[HOT98] H ODGINS J. K., O’B RIEN J. F., T UMBLIN J.:
Perception of human motion with different geometric
models. IEEE Transactions on Visualization and Computer Graphics 4, 4 (October 1998), 307–316.
[HRvdP04] H ARRISON J., R ENSINK R. A., VAN DE
PANNE M.: Obscuring length changes during animated
motion. ACM Transactions on Graphics 23, 3 (Aug.
2004), 569–573.
[KPWH92] K AISER M. K., P ROFFITT D. R., W HELAN
S. M., H ECHT H.: Influence of animation on dynamical
judgments. Journal of Experimental Psychology: Human
Perception and Performance 18, 3 (1992), 669–690.
[MC91] M ACMILLAN N. A., C REELMAN C. D.: Detection Theory: A User’s Guide. Cambridge University
Press, New York, 1991.
[Mic63] M ICHOTTE A.: The perception of causality.
Metheun, London (1963).
[OD01] O’S ULLIVAN C., D INGLIANA J.: Collisions and
perception. ACM Transactions on Graphics 20, 3 (July
2001), 151–168.
[ODGK03] O’S ULLIVAN C., D INGLIANA J., G IANG T.,
K AISER M. K.: Evaluating the visual fidelity of physically based animations. In ACM Transactions on Graphics 2, 3 (2003).
[OHJ00] O ESKER M., H ECHT H., J UNG B.: Psychological evidence for unconscious processing of detail in realtime animation of multiple characters. Journal of Visualization and Computer Animation 11 (2000), 105–112.
[PMM∗ 03] P ELPHREY K. A., M ITCHELL T. V., M C K E OWN M. J., G OLDSTEIN J., A LLISON T., M C C ARTHY
G.: Brain activity evoked by the perception of human
walking: Controlling for meaningful coherent motion.
The Journal of Neuroscience, 23(17):6819-6825 (2003).
[Pro99] P ROFFITT D. R.: Naive physics. In The MIT Encyclopedia of the Cognitive Sciences, Wilson R., Keil F.,
(Eds.). MIT Press, 1999, pp. 577–579.
[PT71] PATTERSON H., T HOMPSON R.: Recovery of
inter-block information when block sizes are unequal.
Biometrika 58, p.545-554 (1971).
[RP03] R EITSMA P. S. A., P OLLARD N. S.: Perceptual
metrics for character animation: Sensitivity to errors in
ballistic motion. ACM Transactions on Graphics 22, 3
(July 2003), 537–542.
[RPE∗ 05] R EN L., PATRICK A., E FROS A. A., H ODGINS
J. K., R EHG J. M.: A data-driven approach to quantifying
natural human motion. ACM Transactions on Graphics
24, 3 (Aug. 2005), 1090–1097.
[Run74] RUNESON S.: Constant velocity - not perceived
as such. Psychological Research, 37(1), pp. 3-23 (1974).
[SW93] S TAPPERS P. J., WALLER P. E.: Using the free
fall of objects under gravity for visual depth estimation.
Bulletin of the Psychonomic Society 31, 2 (1993), 125–
127.
[WB03] WANG J., B ODENHEIMER B.: An evaluation of
a cost metric for selecting transitions between motion segments. In 2003 ACM SIGGRAPH / Eurographics Symposium on Computer Animation (Aug. 2003), pp. 232–238.
[WB04] WANG J., B ODENHEIMER B.: Computing the
duration of motion transitions: an empirical approach. In
2004 ACM SIGGRAPH / Eurographics Symposium on
Computer Animation (July 2004), pp. 335–344.
c 2008 The Author(s)
c 2008 The Eurographics Association and Blackwell Publishing Ltd.
Journal compilation